Overview

This repository contains data and code used in the paper "Predicting Moral Values in Lyrics Through Audio" submitted to the 2025 Content-Based Multimedia Indexing (CBMI) conference.

Human-Annotated Lyrics

• We utilise a dataset of 200 English language song lyrics, annotated with 10 moral values (virtue/vice polarities treated as separate labels) by two bilingual annotators link here.

Audio Excerpts

• The preview URLs are gathered using the utils_get_previews script, and the previews themselves are downloaded using the utils_download_previews script.

Feature Extraction and Dataset Construction

The utils_construct_dataset script extracts features and constructs the dataset, using the audiofeatureextractor class:

• A combination of custom-designed, Essentia, and MELODIA features are extracted.
• The extracted features are saved in a dictionary format, categorized by type for easier filtering or elimination.
• The class includes functionality to convert these dictionaries into Pandas DataFrames, making them ready for use with XGBoost.

Lyrics Moral Predictions from Audio

• The audio-mft.ipynb notebook is used to predict moral foundations based on the extracted audio features.